The limit
A clone cannot beat its source
Cloning reproduces whatever the reference contains, including room reflections and mix noise. From a room microphone, a good stock voice sounds better than the clone.
About
I know how synthetic speech fails.
Years of watching these systems break.
The demo is never the hard part. Real input is.
The tools behind Yazyk are the tools I already work in every day:
- Text-to-speech benchmarking. Measuring what a synthesis engine actually does with real speech: how far behind it runs, where it stutters, where it seams two utterances together badly.
- Voice cloning. Building a speaker's voice from a short reference and learning, repeatedly, that the clone can never be better than the audio it was taken from.
- Lip sync and dubbing. Putting a translated performance back onto a face and a timeline.
- Face swap and generative video. The deepfake side of this field, run locally on my own hardware rather than through somebody's API.
- Voice isolation. Pulling a single speaker out of music, room noise, and a full mix, which is the same problem a live sound desk hands you on the day.
Every one of those systems looks flawless in a curated clip. Finding out where each one falls apart is most of the actual work, and it is the part nobody puts in a demo.
Why it stays on your machine
8 yearsI spent eight years breaking into Fortune 500 companies.
That was the job: get in, reach the data nobody thought was reachable, and write up how. You learn quickly that the weak point is almost never the lock everyone is looking at. It is the convenient thing somebody added later, and the copy of your data sitting somewhere you forgot to ask about.
So Yazyk runs the pipeline on the computer in your room, and your audio does not pass through a server of mine. Not because it is a feature, but because I have spent a career being the person who goes looking for the copy you forgot about.
Voice cloning: a switch, and a default I got wrong.
Yazyk can speak the translation in the speaker's own voice, built from a short reference. Whether that sounds better than a stock voice depends entirely on the feed it is taken from.
The default
On today, and it should not be
The app ships with cloning on. The runbooks disagree: the operator guide says to leave it off unless the feed is a clean board feed, while the spike service command uses --clone auto. The default has no consistent operating instruction, and resolving that is on the list.
The likeness
Someone's voice is theirs
A cloned voice is a copy of a real person. It is built for that event from a reference you supply, and the software deletes the vendor side copy when the run shuts down cleanly. A crash, a slow vendor response at shutdown, or a failed delete can still leave one behind, so the runbook carries a manual check.
Rules I would want applied to me.
Four commitments
- A voice is only cloned from a reference the speaker knowingly gave. Yazyk asks that of you; it cannot enforce it for you.
- The audio pipeline runs on the computer in your booth, not on a server of mine.
- A failure Yazyk recognises appears on the receipt, in the same place as what worked. It recognises a silent input, a dead output, a run that produced no audio, a busy device, and a clone that failed. A fault outside that list is one the receipt will not name, and that list grows every time a real event finds a new one.
- Claims on this site say what is running today and what is not.
The last one is why this page tells you the cloning default is wrong, why nothing is reported back to us yet, and why the vendor side voice can still outlive a run that crashed. Those are the current state, not the intended one.
Who you are dealing with
A small operation, named plainly.
Yazyk is built and operated by Tymeless Technologies LLC. Paid plans are not open yet, so the honest description of the product today is early software running real events and reporting what it measured of them.
Talk about your event